Communications of the ACM
How Computers Can Learn Better
MIT News
A new reinforcement-learning algorithm enables computer systems to find solutions much more efficiently than they did with previous algorithms, researchers say.
Credit: University of Pennsylvania
Massachusetts Institute of Technology (MIT) researchers say they have developed a reinforcement-learning algorithm that enables computer systems to find solutions much more efficiently than previous algorithms.
The researchers also have developed a programming framework designed to make it much easier to set up and run reinforcement-learning experiments.
The software, called RLPy, should enable researchers to more efficiently test new algorithms and compare algorithms' performance on different tasks, says MIT's Alborz Geramifard.
The new algorithm identifies pertinent features in reinforcement-learning tasks by building a data structure that represents different combinations of features. The algorithm then determines which combinations of features dictate a policy's success or failure.
Geramifard says the new algorithm identifies an initial feature on which to base judgments and then looks for complementary features that can refine the initial judgment. "Think of it as like a Lego set," he says. "You can snap one module out and snap another one in its place."
During testing, the researchers say their algorithm evaluated policies and produced more reliable predictions in one-fifth the time of previous algorithms.
From MIT News
View Full Article
Abstracts Copyright © 2013 Information Inc., Bethesda, Maryland, USA
No entries found